Stateless sessions
Each session starts clean, so continuity has to come from somewhere outside the model.
Cross-session coordination for AI coding agents
Stateless AI agents need an environment, not a conversation.
Termite Protocol lets fresh AI agent sessions pick up real repository work through SQLite signals, atomic claiming, compact .birth snapshots, and cross-session field memory.
Environment-first coordination
Mixed-strength model leverage
Audit-backed design
6
production colonies
4
multi-model audits
900+
total commits
96.4%
A-005 observation quality
The field carries the coordination.
Signals persist in SQLite, claims stay atomic, and new sessions read a bounded operational snapshot instead of replaying a long conversation.
.birth ≤ 800 tokensFresh sessions read an arrival snapshot instead of a giant rulebook.Atomic signal claimingNo scheduler bottleneck, no double-assignment, no conversation overhead.
Shepherd EffectOne strong model can seed field patterns that weaker workers follow.
Problem
The session ends. The context disappears. The project does not.
Most AI coding agents are stateless. Every new session risks rediscovering the same structure, repeating the same mistakes, and losing the same design context. Conversation-heavy coordination helps only until context cost, drift, and weaker models start to dominate.
Coordination drift
Chat-based handoffs paraphrase state again and again, which compounds ambiguity and loss.
Weak model failure
Cheaper models often fail at initiating high-quality patterns on their own, especially over multi-turn conversations.
Mechanism
Put coordination in the field.
Termite Protocol moves collaboration into the repository environment itself. Instead of preserving continuity through an ever-growing chat log, it writes operational state directly into the field.
Environment-first coordination
Agents do not need to talk to each other. They sense shared state and continue the work.
- signals in SQLite
- field memory in observations
- arrival hints in
.birth
Bounded context cost
.birth compresses the current colony state into a compact operational snapshot, so fresh sessions do not reload a 28K-token protocol document.
Mixed-strength model leverage
Strong models can seed patterns that weaker workers follow, improving colony-wide output quality without making every worker equally expensive.
60-second smoke test
mkdir termite-demo && cd termite-demo
curl -fsSL https://raw.githubusercontent.com/
billbai-longarena/Termite-Protocol/main/install.sh | bash
./scripts/field-arrive.sh
./scripts/field-pulse.sh
sqlite3 .termite.db "select id,status,title from signals;"Proof
Backed by field data, not just architecture diagrams.
The protocol has been exercised across real colonies and audit packages, including a Shepherd Effect configuration where one strong model and two weaker workers reached 96.4% observation quality.
6production colonies
4multi-model audit experiments
900+total commits
96.4%A-005 observation quality
A-003 · weak-only
35.7%
Weak models can execute the loop, but they struggle to sustain judgment and structure independently.
A-005 · Shepherd Effect
96.4%
One strong model seeds the field, later weaker workers imitate the pattern, and the colony behaves far above weak-only quality.
A-006 · 5-model mix
57%
Scale boosts throughput, but the field still needs mechanisms that prevent dilution and starvation.
Workflow
Fresh sessions become continuous work.
Each new agent reads a bounded snapshot, claims work atomically, executes, and writes observations back into the field for the next session.
1
Arrive
field-arrive.sh computes .birth from the current colony state.
2
Claim
field-claim.sh atomically claims the next unassigned signal.
3
Act
The agent completes work using the field snapshot instead of replaying full chat history.
4
Deposit
field-deposit.sh leaves observations, decisions, and status for the next worker.
Good fit
- multi-agent parallel development
- long-running repositories
- strong + weak model mixes
- audit-heavy engineering teams
- large refactors with many separable tasks
Not for everything
- tiny one-off tasks
- purely exploratory work
- small scripts with no continuity needs
- cases where a single strong interactive agent is simpler
FAQ
What people usually ask first.
How is this different from conversation-based multi-agent tools?
Conversation is not the coordination backbone. The environment is. The repository stores signals, field memory, and recovery hints so the workflow survives session boundaries.
Do I need a separate automation layer to use it?
No. The protocol is installable and usable on its own. Automation layers can sit on top later, but they are optional companions, not prerequisites.
Does this replace strong models?
No. It makes strong models more leverageable and weaker models more usable inside the same field.
Stay in touch
Newsletter, waitlist, and direct contact all live here.
The site is still intentionally lightweight, so the fastest path is email. Pick the path that matches your intent and tell us what you want to build with Termite Protocol.
Newsletter
Get release notes and field learnings.
Subscribe for new releases, audit takeaways, and fresh patterns for cross-session collaboration.
Join newsletterWaitlist
Share your use case and timeline.
Join the waitlist if you want pilot updates, onboarding availability, or a heads-up on the next ecosystem rollout.
Join waitlistContact
Talk directly with Bill.
Use this for partnerships, production questions, research collaboration, or anything that needs a real conversation.
Email Bill
Prefer a plain email thread? Write to bill.bai@long-arena.com and include your repository, team size, or rollout questions.
Test where stateless agents stop feeling stateless.
Run the smoke test, inspect the audit materials, and stress the assumptions in a real repository. The fastest path is the repo itself; the strongest proof is still skeptical field use.