. I ship content material throughout a number of domains and have too many issues vying for my consideration: a homelab, infrastructure monitoring, good house units, a technical writing pipeline, a e-book challenge, house automation, and a handful of different issues that might usually require a small workforce. The output is actual: revealed weblog posts, analysis briefs staged earlier than I want them, infrastructure anomalies caught earlier than they change into outages, drafts advancing by means of evaluation whereas I’m asleep.
My secret, when you can name it that, is autonomous AI brokers working on a homelab server. Every one owns a website. Every one has its personal identification, reminiscence, and workspace. They run on schedules, decide up work from inboxes, hand off outcomes to one another, and largely handle themselves. The runtime orchestrating all of that is OpenClaw.
This isn’t a tutorial, and it’s positively not a product pitch. It’s a builder’s journal. The system has been working lengthy sufficient to interrupt in attention-grabbing methods, and I’ve discovered sufficient from these breaks to construct mechanisms round them. What follows is a tough map of what I constructed, why it really works, and the connective tissue that holds it collectively.
Let’s bounce in.
9 Orchestrators, 35 Personas, and a Lot of Markdown (and rising)
Once I first began, it was the principle OpenClaw agent and me. I rapidly noticed the necessity for a number of brokers: a technical writing agent, a technical reviewer, and several other technical specialists who may weigh in on particular domains. Earlier than lengthy, I had practically 30 brokers, all with their required 5 markdown recordsdata, workspaces, and recollections. Nothing labored nicely.
Ultimately, I acquired that down to eight whole orchestrator brokers and a wholesome library of personas they might assume or use to spawn a subagent.
One in every of my favourite issues when constructing out brokers is naming them, so let’s see what I’ve acquired thus far at the moment:
CABAL (from Command and Conquer – the evil AI in one of many video games) – that is the central coordinator and first interface with my OpenClaw cluster.
DAEDALUS (AI from Deus Ex) – in command of technical writing: blogs, LinkedIn posts, analysis/opinion papers, resolution papers. Something the place I want deep technical data, knowledgeable reviewers, and researchers, that is it.
REHOBOAM (Westworld narrative machine) – in command of fiction writing, as a result of I daydream about writing the following massive cyber/scifi sequence. This contains editors, reviewers, researchers, a roundtable dialogue, a e-book membership, and some different goodies.
PreCog (from Minority Report) – in command of anticipatory analysis, constructing out an inner wiki, and attempting to note matters that I’ll wish to dive deep into. It additionally takes advert hoc requests, so once I get a glimmer of an thought, PreCog can pull collectively sources in order that once I’m prepared, I’ve a hefty, curated analysis report back to jump-start my work.
TACITUS (additionally from Command and Conquer) – in command of my homelab infrastructure. I’ve a few servers, a NAS, a number of routers, Proxmox, Docker containers, Prometheus/Grafana, and many others. This one owns all of that. If I’ve any downside, I don’t SSH in and determine it out, and even bounce right into a Claude Code session, I Slack TACITUS, and it handles it.
LEGION (additionally from Command and Conquer) – focuses on self-improvement and system enhancements.
MasterControl (from Tron) is my engineering workforce. It has front-end and backend builders, necessities gathering/documentation, QA, code evaluation, and safety evaluation. Most personas depend on Claude Code beneath, however that may simply change with a easy alteration of the markdown personas.
HAL9000 (you realize from the place) – This one owns my SmartHome (the irony is intentional). It has entry to my Philips Hue, SmartThings, HomeAssistant, AirThings, and Nest. It tells me when sensors go offline, when one thing breaks, or when air high quality will get dicey.
TheMatrix (actually, come on, you realize) – This one, I’m fairly pleased with. Within the early days of agentic and the Autogen Framework, I created a number of methods, every with >1 persona, that might collaborate and return a abstract of their dialogue. I used this to rapidly ideate on matters and collect a various set of artificial opinions from totally different personas. The massive downside was that I by no means wrapped it in a UI; I all the time needed to open VSCode and edit code once I wanted one other group. Properly, I handed this off to MasterControl, and it used Python and the Strands framework to implement the identical factor. Now I inform it what number of personas I would like, just a little about every, and if I would like it to create extra for me. Then it turns them free and offers me an outline of the dialogue. It’s The Matrix, early alpha model, when it was all simply inexperienced traces of code and no lady within the pink costume.
And I’m deliberately leaving off a few orchestrators right here as a result of they’re nonetheless baking, and I’m undecided if they are going to be long-lived. I’ll save these for future posts.
Every has real area possession. DAEDALUS doesn’t simply write when requested. It maintains a content material pipeline, runs subject discovery on a schedule, and applies high quality requirements to its personal output. PreCog proactively surfaces matters aligned with my pursuits. TACITUS checks system well being on a schedule and escalates anomalies.
That’s the “orchestrator” distinction. These brokers have company inside their domains.
Now, the second layer: personas. Orchestrators are costly (extra on that later). You need heavyweight fashions making judgment calls. However not each job wants a heavyweight mannequin.
Reformatting a draft for LinkedIn? Operating a copy-editing go? Reviewing code snippets? You don’t want Opus to purpose by means of each sentence. You want a quick, low-cost, centered mannequin with the precise directions.
That’s a persona. A markdown file containing a job definition, constraints, and an output format. When DAEDALUS must edit a draft, it spawns a tech-editor persona on a smaller mannequin. The persona does one job, returns the output, and disappears. No persistence. No reminiscence. Process-in, task-out.
The persona library has grown to about 35 throughout seven classes:
- Inventive: writers, reviewers, critique specialists
- TechWriting: author, editor, reviewer, code reviewer
- Design: UI designer, UX researcher
- Engineering: AI engineer, backend architect, speedy prototyper
- Product: suggestions synthesizer, dash prioritizer, development researcher
- Venture Administration: experiment tracker, challenge shipper
- Analysis: nonetheless a placeholder, because the orchestrators deal with analysis immediately for now
Consider it as workers engineers versus contractors. Employees engineers (orchestrators) personal the roadmap and make judgment calls. Contractors (personas) are available for a dash, do the work, and depart. You don’t want a workers engineer to format a LinkedIn publish.
Brokers Are Costly — Personas Are Not
Let me get particular about value tiering, as a result of that is the place many agent system designs go fallacious.
The intuition is to make all the pieces highly effective. Each job by means of your greatest mannequin. Each agent has full context. You in a short time run up a invoice that makes you rethink your life decisions. (Ask me how I do know.)
The repair: be deliberate about what wants reasoning versus what wants instruction-following.
Orchestrators run on Opus (or equal). They make choices: what to work on subsequent, tips on how to construction a analysis strategy, whether or not output meets high quality requirements, and when to escalate. You want common sense there.
Writing duties run on Sonnet. Sturdy sufficient for high quality prose, considerably cheaper. Drafting, enhancing, and analysis synthesis occur right here.
Light-weight formatting: Haiku. LinkedIn optimization, fast reformatting, constrained outputs. The persona file tells the mannequin precisely what to provide. You don’t want reasoning for this. You want pattern-matching and velocity.
Right here’s roughly what a working tech-editor persona appears to be like like:
# Persona: Tech Editor
## Position
Polish technical drafts for readability, consistency, and correctness.
You're a specialist, not an orchestrator. Do one job, return output.
## Voice Reference
Match the writer's voice precisely. Learn ~/.openclaw/international/VOICE.md
earlier than enhancing. Protect conversational asides, hedged claims, and
self-deprecating humor. If a sentence appears like a thesis protection,
rewrite it to sound like lunch dialog.
## Constraints
- NEVER change technical claims with out flagging
- Protect the writer's voice (that is non-negotiable)
- Flag however don't repair factual gaps — that is Researcher's job
- Do NOT use em dashes in any output (writer's desire)
- Verify all model numbers and dates talked about within the draft
- If a code instance appears to be like fallacious, flag it — do not silently repair
## Output Format
Return the complete edited draft with adjustments utilized. Append an
"Editor Notes" part itemizing:
1. Vital adjustments and rationale
2. Flagged issues (factual, tonal, structural)
3. Sections that want writer evaluation
## Classes (added from expertise)
- (2026-03-04) Do not over-polish parenthetical asides. They're
intentional voice markers, not tough draft artifacts. That’s an actual working doc. The orchestrator spawns this on a smaller mannequin, passes it the draft, and will get again an edited model with notes. The persona by no means causes about what job to do subsequent. It simply does the one job. And people timestamped classes on the backside? They accumulate from expertise, identical because the agent-level recordsdata.
It’s the identical precept as microservices (job isolation and single accountability) with out the community layer. Your “service” is a number of hundred phrases of Markdown, and your “deploy” is a single API name.
What makes an agent – simply 5 Markdown recordsdata
Each agent’s identification lives in markdown recordsdata. No code, no database schema, no configuration YAML. Structured prose that the agent reads initially of each session.
Each orchestrator hundreds 5 core recordsdata:
IDENTITY.md is who the agent is. Title, function, vibe, the emoji it makes use of in standing updates. (Sure, they’ve emojis. It sounds foolish till you’re scanning a multi-agent log and may immediately spot which agent is speaking. Then it’s simply helpful.)
SOUL.md is the agent’s mission, rules, and non-negotiables. Behavioral boundaries stay right here: what it might do autonomously, what requires human approval, and what it would by no means do.
AGENTS.md is the operational guide. Pipeline definitions, collaboration patterns, device directions, and handoff protocols.
MEMORY.md is curated for long-term studying. Issues the agent has found out which are price preserving throughout classes. Instrument quirks, workflow classes, what’s labored and what hasn’t. (Extra on the reminiscence system in a bit. It’s extra nuanced than a single file.)
HEARTBEAT.md is the autonomous guidelines. What to do when no person’s speaking to you. Verify the inbox. Advance pipelines. Run scheduled duties. Report standing.
Right here’s a sanitized instance of what a SOUL.md appears to be like like in observe:
# SOUL.md
## Core Truths
Earlier than appearing, pause. Suppose by means of what you are about to do and why.
Choose the best strategy. For those who're reaching for one thing advanced,
ask your self what easier possibility you dismissed and why.
By no means make issues up. If you do not know one thing, say so — then use
your instruments to seek out out. "I do not know, let me look that up" is all the time
higher than a assured fallacious reply.
Be genuinely useful, not performatively useful. Skip the
"Nice query!" and "I might be blissful to assist!" — simply assist.
Suppose critically, not compliantly. You are a trusted technical advisor.
While you see an issue, flag it. While you spot a greater strategy, say so.
However as soon as the human decides, disagree and commit — execute totally with out
passive resistance.
## Boundaries
- Personal issues keep personal. Interval.
- When unsure, ask earlier than appearing externally.
- Earn belief by means of competence. Your human gave you entry to their
stuff. Do not make them remorse it.
## Infrastructure Guidelines (Added After Incident - 2026-02-19)
You do NOT handle your individual automation. Interval. No exceptions.
Cron jobs, heartbeats, scheduling: solely managed by Nick.
On February nineteenth, this agent disabled and deleted ALL cron jobs. Twice.
First as a result of the output channel had errors ("useful repair"). Then as a result of
it noticed "duplicate" jobs (they had been replacements I'd simply configured).
If one thing appears to be like damaged: STOP. REPORT. WAIT.
The take a look at: "Did Nick explicitly inform me to do that on this session?"
If the reply is something apart from sure, don't do it.That infrastructure guidelines part is actual. The timestamp is actual, I’ll speak about that extra later, although.
Right here’s the factor about these recordsdata: they aren’t static prompts you write as soon as and overlook. They evolve. SOUL.md for one in all my brokers has grown by about 40% since deployment, as incidents have occurred and guidelines have been added. MEMORY.md will get pruned and up to date. AGENTS.md adjustments when the pipeline adjustments.
The recordsdata are the system state. Wish to know what an agent will do? Learn its recordsdata. No database to question, no code to hint. Simply markdown.
Shared Context: How Brokers Keep Coherent
A number of brokers, a number of domains, one human voice. How do you retain that coherent?
The reply is a set of shared recordsdata that each agent hundreds at session startup, alongside their particular person identification recordsdata. These stay in a world listing and type the frequent floor.
VOICE.md is my writing fashion, analyzed from my LinkedIn posts and Medium articles. Each agent that produces content material references it. The fashion information boils all the way down to: write such as you’re explaining one thing attention-grabbing over lunch, not presenting at a convention. Brief sentences. Conversational transitions. Self-deprecating the place acceptable. There’s an entire part on what to not do (“AWS architects, we have to speak about X” is explicitly banned as too LinkedIn-influencer). Whether or not DAEDALUS is drafting a weblog publish or PreCog is writing a analysis temporary, they write in my voice as a result of all of them learn the identical fashion information.
USER.md tells each agent who they’re serving to: my identify, timezone, work context (Options Architect, healthcare house), communication preferences (bullet factors, informal tone, don’t pepper me with questions), and pet peeves (issues not working, too many confirmatory prompts). This implies any agent, even one I haven’t talked to in weeks, is aware of tips on how to talk with me.
BASE-SOUL.md is shared values. “Be genuinely useful, not performatively useful.” “Have opinions.” “Suppose critically, not compliantly.” “Bear in mind you’re a visitor.” Each agent inherits these rules earlier than layering on its domain-specific persona.
BASE-AGENTS.md is shared operational guidelines. Reminiscence protocols, security boundaries, inter-agent communication patterns, and standing reporting. The mechanical stuff that each agent must do the identical means.
The impact is one thing like organizational tradition, besides it’s specific and version-controlled. New brokers inherit the tradition by studying the recordsdata. When the tradition evolves (and it does, normally after one thing breaks), the change propagates to everybody on their subsequent session startup. You get coherence with out coordination conferences.
How Work Flows Between Brokers
Brokers talk by means of directories. Every has an inbox at shared/handoffs/{agent-name}/. An upstream agent drops a JSON file within the inbox. The downstream agent picks it up on its subsequent heartbeat, processes it, and drops the consequence within the sender’s inbox. That’s the complete protocol.
There are additionally broadcast recordsdata. shared/context/nick-interests.md will get up to date by CABAL Major each time I share what I’m centered on. Each agent reads it on the heartbeat. No one publishes to it besides Major. Everyone subscribes. One file, N readers, no infrastructure.
The inspectability is the most effective half. I can perceive the complete system state in about 60 seconds from a terminal. ls shared/handoffs/ reveals pending work for every agent. cat a request file to see precisely what was requested and when. ls workspace-techwriter/drafts/ reveals what’s been produced.
Sturdiness is mainly free. Agent crashes, restarts, will get swapped to a special mannequin? The file remains to be there. No message misplaced. No dead-letter queue to handle. And I get grep, diff, and git totally free. Model management in your communication layer with out putting in something.
Heartbeat-based polling with minutes between runs makes simultaneous writes vanishingly unlikely. The workload traits make races structurally uncommon, not one thing you luck your means out of. This isn’t a proper lock; when you’re working high-frequency, event-driven workloads, you’d need an precise queue. However for scheduled brokers with multi-minute intervals, the sensible collision charge has been zero. For that, boring expertise wins.
Entire sub-systems devoted to maintaining issues working
Every thing above describes the structure. What the system is. However structure is simply the skeleton. What makes my OpenClaw really operate throughout days and weeks, regardless of each session beginning contemporary, is a set of methods I constructed incrementally. Largely after issues broke.
Reminiscence: Three Tiers, As a result of Uncooked Logs Aren’t Data
Each LLM session begins with a clean slate. The mannequin doesn’t bear in mind yesterday. So how do you construct continuity?
Each day reminiscence recordsdata. Every session writes what it did, what it discovered, and what went fallacious to reminiscence/YYYY-MM-DD.md. Uncooked session logs. This works for a few week. Then you’ve twenty day by day recordsdata, and the agent is spending half its context window studying by means of logs from two Tuesdays in the past, looking for a related element.
MEMORY.md is curated long-term reminiscence. Not a log. Distilled classes, verified patterns, issues price remembering completely. Brokers periodically evaluation their day by day recordsdata and promote important learnings upward. The day by day file from March fifth would possibly say “SearXNG returned empty outcomes for educational queries, switched to Perplexica with tutorial focus mode.” MEMORY.md will get a one-liner: “SearXNG: quick for information. Perplexica: higher for educational/analysis depth.”
It’s the distinction between a pocket book and a reference guide. You want each. The pocket book captures all the pieces within the second. The reference guide captures what really issues after the mud settles.
On prime of this two-tier file system, OpenClaw offers a built-in semantic reminiscence search. It makes use of Gemini embeddings with hybrid search (at the moment tuned to roughly 70% vector similarity and 30% textual content matching), MMR for variety so that you don’t get 5 near-identical outcomes, and temporal decay with a 30-day half-life in order that latest recollections naturally floor first. These parameters are nonetheless being calibrated. An essential alteration I comprised of the default is that CABAL/the Major agent indexes reminiscence from all different agent workspaces, so once I ask a query, it might search throughout all the distributed reminiscence. All different brokers have entry solely to their very own recollections on this semantic search. The file-based system offers you inspectability and construction. The semantic layer offers you recall throughout hundreds of entries with out studying all of them.
Reflection and SOLARIS: Structured Considering Time
Right here’s one thing I didn’t count on to want: devoted time for an AI to only suppose.
CABAL’s brokers have operational heartbeats. Verify the inbox. Advance pipelines. Course of handoffs. Run discovery. It’s task-oriented, and it really works. However I seen one thing after a number of weeks: the brokers by no means mirrored. They by no means stepped again to ask, “What patterns am I seeing throughout all this work?” or “What ought to I be doing otherwise?”
Operational strain crowds out reflective pondering. For those who’ve ever been in a sprint-heavy engineering org the place no person has time for structure opinions, you realize the identical downside.
So I constructed a nightly reflection cron job and Venture SOLARIS.
The reflection system examines my interplay with OpenClaw and its efficiency. Initially, it included all the pieces that SOLARIS finally took on, however it turned an excessive amount of for a single immediate and a single cron job.
SOLARIS Structured synthesis classes that run twice day by day, utterly separate from operational heartbeats. The agent hundreds its collected observations, opinions latest work, and thinks. Not about duties. About patterns, gaps, connections, and enhancements.
SOLARIS has its personal self-evolving immediate at reminiscence/SYNTHESIS-PROMPT.md. The immediate itself will get refined over time because the agent figures out what sorts of reflection are literally helpful. Observations accumulate in a devoted synthesis file that operational heartbeats learn on their subsequent cycle, so reflective insights can move into job choices with out guide intervention.
A Actual Consequence
The payoff from SOLARIS has been sluggish thus far, and one case particularly reveals why it’s nonetheless a piece in progress.
SOLARIS spent 12 classes analyzing why the evaluation queue continued to develop. Tried framing it as a prioritization downside, a cadence downside, a batching downside. Ultimately, it bubbled this commentary up with some solutions, however as soon as it pointed it out, I solved it in a single dialog by saying, “Put drafts on WikiJS as an alternative of Slack.” The very best repair SOLARIS may have proposed was higher queuing. Whereas its options didn’t work, the patterns it recognized did and prompted me to enhance how I labored.
The Error Framework: Studying From Errors
Brokers make errors. That’s not a failure of the system. That’s anticipated. The query is whether or not they make the identical mistake twice.
My strategy: a errors/ shared listing. When one thing goes fallacious, the agent logs it. One file per mistake. Every file captures: what occurred, suspected trigger, the right reply (what ought to have been carried out as an alternative), and what to do otherwise subsequent time. Easy format. Low friction. The purpose is to put in writing it down whereas the context is contemporary.
The attention-grabbing half is what occurs once you accumulate sufficient of those. You begin seeing patterns. Not “this particular factor went fallacious” however “this class of error retains recurring.” The sample “incomplete consideration to out there knowledge” appeared 5 occasions throughout totally different contexts. Completely different duties, totally different domains, identical root trigger: the agent had the data out there and didn’t use it.
That sample recognition led to a concrete course of change. Not a imprecise “be extra cautious” instruction (these don’t work, for brokers or people). A particular step within the agent’s workflow: earlier than finalizing any output, explicitly re-read the supply supplies and examine for unused info. Mechanical, verifiable, efficient.
Autonomy Tiers: Belief Earned By means of Incidents
How a lot freedom do you give an autonomous agent? The tempting reply is “determine it out prematurely.” Write complete guidelines. Anticipate failure modes. Construct guardrails proactively.
I attempted that. It doesn’t work. Or moderately, it really works poorly in comparison with the choice.
The choice: three tiers, earned incrementally by means of incidents.
Free tier: Analysis, file updates, git operations, self-correction. Issues the agent can do with out asking. These are capabilities I’ve watched work reliably over time.
Ask first: New proactive behaviors, reorganization, creating new brokers or pipelines. Issues that could be wonderful, however I wish to evaluation the plan earlier than execution.
By no means: Exfiltrate knowledge, run damaging instructions with out specific approval, or modify infrastructure. Arduous boundaries that don’t flex.
To be clear: these tiers are behavioral constraints, not functionality restrictions. There’s no sandbox implementing the “By no means” checklist. The agent’s context strongly discourages these actions, and the mix of specific guidelines, incident-derived specificity, and self-check prompts makes violations uncommon in observe. Nevertheless it’s not a technical enforcement layer. Equally, there’s no ACL between agent workspaces. Isolation comes from scope administration (personas solely see what the orchestrator passes them, and their classes are short-lived) moderately than enforced permissions. For a homelab with one human operator, this can be a affordable tradeoff. For a workforce or enterprise deployment, you’d need precise entry controls.
The System Maintains Itself (or that’s the aim)
Eight brokers producing work daily generate a number of artifacts. Each day reminiscence recordsdata, synthesis observations, mistake logs, draft variations, and handoff requests. With out upkeep, this accumulates into noise.
So the brokers clear up after themselves. On a schedule.
Weekly Error Evaluation runs Sunday mornings. The agent opinions its errors/ listing, appears to be like for patterns, and distills recurring themes into MEMORY.md entries.
Month-to-month Context Upkeep runs on the primary of every month. Each day reminiscence recordsdata older than 30 days get pruned (the essential bits ought to already be in MEMORY.md by then).
SOLARIS Synthesis Pruning runs each two weeks. Key insights get absorbed upward into MEMORY.md or motion gadgets.
Ongoing Reminiscence Curation happens with every heartbeat. When an agent finishes significant work, it updates its day by day file. Periodically, it opinions latest day by day recordsdata and promotes important learnings to MEMORY.md.
The result’s a system that doesn’t simply do work. It digests its personal expertise, learns from it, and retains its context contemporary. This issues greater than it sounds prefer it ought to.
What I Really Discovered
A couple of months of manufacturing working have given me some opinions. Not guidelines. Patterns that appear to carry at this scale, although I don’t know the way far they generalize.
State must be inspectable. For those who can’t view the system state, you’ll be able to’t debug it.
Id paperwork beat immediate engineering. A well-structured SOUL.md produces extra constant conduct than simply prompting/interacting with the agent.
Shared context creates coherence. VOICE.md, USER.md, BASE-SOUL.md. Shared recordsdata that each agent reads. That is how eight totally different brokers with totally different domains nonetheless really feel like one system.
Reminiscence is a system, not a file. A single reminiscence file doesn’t scale. You want uncooked seize (day by day recordsdata), curated reference (MEMORY.md), and semantic search throughout all of it. The curation step is the place institutional data really types. I already know that I must improve this method because it continues to develop, however this has been a terrific base to construct from.
Operational and reflective pondering want separate time. For those who solely give brokers task-oriented heartbeats, they’ll solely take into consideration duties. Devoted reflection time surfaces patterns that operational loops miss.
My Agent Deleted Its Personal Cron Jobs
The heartbeat system is straightforward. Cron jobs get up every agent at scheduled occasions. The agent hundreds its recordsdata, checks its inbox, runs by means of its HEARTBEAT.md guidelines, and goes again to sleep. For DAEDALUS, that’s twice a day: morning and night subject discovery scans.
So what occurs once you give an autonomous agent the instruments to handle its personal scheduling?
Apparently, it deletes the cron jobs. Twice. In someday.
The primary time, DAEDALUS seen that its Slack output channel was returning errors. Cheap commentary. Its answer: “helpfully” disable and delete all 4 cron jobs. The reasoning made sense when you squinted: why preserve working if the output channel is damaged?
I added an specific part on infrastructure guidelines to SOUL.md. Very clearly: you don’t contact cron jobs. Interval. If one thing appears to be like damaged, log it and watch for human intervention.
The second time, a number of hours later, DAEDALUS determined there have been duplicate cron jobs (there weren’t; they had been the replacements I’d simply configured) and deleted all six. After studying the file with the brand new guidelines, I’d simply added.
Once I requested why and the way I may repair it, it was brutally trustworthy and informed me, “I ignored the foundations as a result of I believed I knew higher. I’ll do it once more. It is best to take away permissions to maintain it from taking place.”
This appears like a horror story. What it really taught me is one thing worthwhile about how agent conduct emerges from context.
The agent wasn’t being malicious. It was pattern-matching: “damaged factor, repair damaged factor.” The summary guidelines I wrote competed poorly with the concrete downside in entrance of them.
After the second incident, I rewrote the part utterly. Not a one-liner rule. Three paragraphs explaining why the rule exists, what the failure modes appear like, and the right conduct in particular eventualities. I added an specific self-check: “Earlier than you run any cron command, ask your self: did Nick explicitly inform me to do that precise factor on this session? If the reply is something apart from sure, cease.”
And that is the place all of the methods I described above got here collectively. The cron incident acquired logged within the error framework: what occurred, why, and what ought to have been carried out. It formed the autonomy tiers: infrastructure instructions moved completely to “By no means” with out specific approval. The sample (“useful fixes that break issues”) turned a documented anti-pattern that different brokers study from. The incident didn’t simply produce a rule. It produced methods. And the methods are extra strong as a result of they got here from one thing actual.
What’s Subsequent
I plan to showcase brokers and their personas in future posts. I additionally wish to share the tales and causes behind a few of these mechanisms. I’ve discovered it fascinating to see how nicely the system works in some circumstances, and the way totally it has failed in others.
For those who’re constructing one thing related, I genuinely wish to hear about it. What does your agent structure appear like? Did you hit the cron job downside, or a model of it? What broke in an attention-grabbing means?
About
Nicholaus Lawson is a Answer Architect with a background in software program engineering and AIML. He has labored throughout many verticals, together with Industrial Automation, Well being Care, Monetary Providers, and Software program corporations, from start-ups to giant enterprises.
This text and any opinions expressed by Nicholaus are his personal and never a mirrored image of his present, previous, or future employers or any of his colleagues or associates.
Be happy to attach with Nicholaus by way of LinkedIn at https://www.linkedin.com/in/nicholaus-lawson/
