Enforced process, deterministic memory
Everyone ships workflows now.Nobody enforces them.
We operate domain solutions with your team, and calibrate custom harnesses for domains beyond them. A process the agent cannot skip; a memory that knows what it knows.
a5c.ai operates AI-agent solutions with your team, and calibrates custom harnesses for domains beyond the ones we package. Two pillars carry every engagement: a process the agent cannot skip, and a memory that knows what it knows.
Custom harnesses are operational with clients today across multiple domains. The enforced process ships now; deterministic memory is in early access.
01How it works
Two pillars keep a run on target.
A harness is everything above the model: the coding agent (Claude Code, Codex, Cursor, and 9 more), its skills, tools, process, and memory.
The run converges to what you intended.
The familiar failure
The context fills, constraints drop, and the agent drifts off target.
The process is code, held outside the model's context.
Every step ends at a hard stop; the process decides what's next.
Run state lives on disk; a reset can't erase the plan.
You stop babysitting the run.
The agent knows what it did, at every step.
The quieter failure
A migration reported done, the agent's word the only evidence.
Every fact is signed and append-only; same facts, same answer.
Disagreement is a typed conflict, never silently resolved.
Any past belief is a query, with author and timestamp.
You stop taking the agent's word for it.
02The contract
Five rules the run cannot break.
- 01Gates block until their criteria are met, with evidence.
- 02Budgets cap time, cost, and tool calls by policy.
- 03Breakpoints stop the run until a human approves.
- 04The journal is append-only.
- 05Every run replays, deterministically.
Inside it, the model does what models are for; the run can only do what the process code permits.
03The terminal
A gate refuses. A dead session resumes. Memory answers.
a gate refuses
Done is proven, or the run does not move.
gate.checkedgate.checked verify-rows: 2 of 3 met
gate.refusedgate.refused progression blocked; revise dispatched
task.completedtask.completed revise-migration: output recorded
gate.passedgate.passed 3 of 3 met, with evidence; next step unlocked
a dead session resumes
The journal, not the context window, is the run.
session.endedsession.ended harness stopped mid-run; context gone
journal.readjournal.read ~/.a5c/runs/migrate-billing/ (append-only, on disk)
state.rebuiltstate.rebuilt plan, budgets, target restored
run.resumedrun.resumed from the step after the last stop
memory answers
The past is a query, not a recollection.
fact.recordedfact.recorded rows-moved: 41,209 of 41,209, signed and appended
conflict.surfacedconflict.surfaced agent summary disagrees with the count; typed, nothing overwritten
memory.asofmemory.asof what was believed at 09:14 answers exactly
04The argument
First steps stop at convenience.
We introduced process as code more than a year ago. Anthropic now ships workflows in Claude Code; the direction is right.
But per-vendor workflow features stay tied to their own harness, and enforcement is not their default. Our enforced process runs the same on all 12 supported harnesses, Anthropic’s included. And workflow features leave the other half untouched: what an agent knows still lives in a mutable store.
A more capable model closes neither half; the fix is architectural.
05Reliability
What reliability means here.
- 01Bounded behavior
- A run touches only what you allowed.
- 02Convergent outcomes
- Work meets your bar or a person is called in.
- 03Evidenced completion
- Done must be proven with evidence.
- 04Replayable history
- Any decision replays in a postmortem, and what the run learned is answerable as of any date.
- 05Directed questions
- A question goes to the person who owns the decision, and the answer is remembered, never re-asked.
06Solutions
Packaged solutions, or your domain.
An operated solution for your domain, run with your team; or a custom harness trained against your evals: tasks with known-good outcomes.
Security, the flagship research lab.
07Under the hood
Many pillars, running as one stack.
Babysitter
The process pillar, as code: gates, budgets, breakpoints, journal, enforced at runtime.
Read more
KIP
The memory pillar, as a ledger: signed facts, deterministic projection, typed conflicts, asOf answers. Early access on request.
Read more
Adapters carries one process across all 12 supported harnesses, no coding-agent lock-in; Kradle and Atlas complete the stack.
The whole map
Every agent runs on the same eleven layers. See them as one building.
The agent stack, from model to surface, drawn live from our Atlas graph and walkable floor by floor.
Walk the stack
08The moat
What compounds with every run.
The code is copyable on purpose. The record of enforced runs is not.
Talk to us.
A demo is a live run against a target you choose.
Or emailhello@a5c.ai
The intro call is 30 minutes and free.