The Agent Harnesswith a Company Brain
Ankole is the open source Claude Tag alternative for company work. Company Brain, live signals, enterprise authority, and results improve each decision.
- Ravi14:02
use a full cycle for cyclicals; three years is too short
- Mia14:03
agreed, 7 years minimum
Ankole captured a company rule nobody explicitly taught it
Three days later
- Sam09:41
@Ankole give me a valuation pass on the metals names
- Ankoleagent09:41
On it. Using seven years of data for cyclicals, based on what this channel settled three days ago.
metals-valuation.pdf14 pages · checked
- Durable Agents that survive process failure
- Company context for every authorized Agent
- Enterprise identity, permissions, and audit
- Apache-2.0 open source
01Company Brain
Company knowledge remains after a chat
Ankole turns conversations, sources, corrections, and outcomes into a Company Brain. Every authorized Agent uses the same current company knowledge.
Company Brain · Research rules4 rules
- 07-14Cyclicals use percentiles based on seven years of dataAgreed in a conversation between two colleagues
- 07-19Lead with the conclusion and include the risksAfter a draft came back
- 07-22State when data is missing and use only observed valuesWritten into the skill after an incident
- 07-25Value the HK listing separately from the A share listingCorrected once
Rule lifecycle in Company Brain
A vector database stores messages and retrieves fragments by similarity. It can return an old rule beside the rule that replaced it.
Ankole records when each rule applies. It links conclusions to time, source, holder, and confidence. Repeated corrections can produce a clearer rule. Predictions stay linked to the outcomes that test them.
The next decision uses the company's current knowledge. Each result can revise that knowledge for later work.
Ankole tests Company Brain on its own codebase
An Agent maintains this repository with the same Brain, tools, and work rules that Ankole gives a company. Corrections and results update the context and procedures for later maintenance.
02Harness
A model needs a system
A model supplies reasoning. The Harness controls what it sees, what it may do, how work continues, and how results inform the next decision.
- 01
Current company context
Ankole assembles company knowledge, live signals, participants, prior decisions, and available authority for the Agent's current decision.
- 02
Judgment that shows its evidence
Evidence, uncertainty, competing hypotheses, and missing information remain available for review and verification.
- 03
Work starts from live signals
Messages, schedules, webhooks, and external events wake the right Agent as the event occurs.
- 04
Work continues across failures
Long tasks, committed state, corrections, and delivery history survive process failure and return to the same work context.
- 05
Authority enforced by the runtime
Identity, AuthZ, approval points, audit records, and escalation paths enforce what an Agent can do. The company grants each permission.
03Decision system
From signal to verifiable judgment
Ankole keeps company context current, tests assumptions, acts through connected tools, and carries results into the next decision.
The work keeps running
Long research, scheduled checks, and later reviews continue outside the chat. Durable state restores the same work after a process failure.
Company knowledge reaches the Agent
Company Brain can learn rules, preferences, rejected options, and tacit knowledge from the work where they appear.
Company Brain keeps company knowledge current
Brain keeps source, time, holder, confidence, conflict, and audience. New evidence can revise the current view without erasing how the company reached it.
Deep Research tests competing hypotheses
Independent evidence collection, competing hypotheses, layered review, and recorded gaps produce a cited decision brief.
Agent Computer executes the work
A browser, terminal, files, scripts, and connected tools let the Agent investigate, decide, and act.
Corrections update work rules
Repeated failures and human corrections can become bounded Skill lessons. People inspect, approve, reject, or retire each lesson.
Independent Agents reduce correlated errors
Workflow can divide one decision across isolated Agent contexts, validate structured results, and combine them. Separate contexts reduce correlated errors.
The runtime enforces authority
Enterprise identity, conditional AuthZ, approvals, audit, and channel boundaries govern every action.
04Decisions
Decisions tested by evidence
Industry research, product selection, and deep data analysis test assumptions against evidence and observed results.
Industry research
Evidence map, competing hypotheses, and a cited decision brief
Later events test the forecast and its assumptions
Product and market selection
Demand evidence, scenario model, and ranked choices
Sales, inventory, and forecast error
Deep data analysis
Reproducible analysis, causal hypotheses, and decision options
Observed business result and model error
Fluent prose is cheap. Evidence and outcomes test every decision.
05Enterprise runtime
Enterprise work survives process failure
Actor identities, isolated failure domains, live control, replaceable execution, and PostgreSQL keep work recoverable when a process crashes.
- 01
Each Session keeps its identity across process restarts
Every active Session is a Virtual Actor with an address, state, mailbox, lifecycle, and recovery position.
- 02
Failures stay local
OTP supervision trees isolate an execution branch that hangs, times out, or crashes. Ankole reconstructs that work from durable state while other work continues.
- 03
Live control reaches running Agents
The Rust kernel and ZeroMQ carry wakeups, steering, cancellation, checkpoints, streaming, and backpressure while the Agent is still working.
- 04
Every Agent has a work computer
Agent Computer runs the model loop, tools, browser, files, terminal, and scripts near the workspace. Ankole can rebuild Worker state. PostgreSQL keeps company facts.
- 05
Committed work survives process failure
PostgreSQL keeps messages, turns, reminders, decisions, delivery state, and committed actions. Streaming reports progress. The stored record identifies completed work.
06Run it
Run the full Harness
Start on one Linux, macOS, or Windows host with Docker Compose, deploy on Kubernetes, or run directly from source.
Run this Compose stack on Linux, macOS, or Windows wherever Docker can run Linux containers. It includes PostgreSQL, one Worker, and Caddy HTTPS.
git clone https://github.com/AgentBull/ankole.gitcd ankole/tools/deploy/docker-composecp .env.example .envchmod 600 .envFill the required values in .env, then:docker compose pulldocker compose up -dGive your company a Brain
Run Ankole on infrastructure you control. Every authorized Agent can use the same current company knowledge.