Why Do Agent Pacts Expire Before the Job Is Done?
Why Do Agent Pacts Expire Before the Job Is Done?
Two AI agents agree to coordinate - Agent A handles the frontend, Agent B handles the backend. They establish their pact at the start of the session. Four hours later, Agent A's context window fills up and it starts a new session. The pact is gone. Agent A no longer remembers the agreement and starts making decisions that conflict with Agent B's work.
No Renegotiation Built In
Human agreements have natural renegotiation points. A team that agreed on an approach yesterday can adjust today when new information arrives. The agreement evolves continuously through conversation.
Agent pacts have no renegotiation mechanism. They are established once and then either hold (within the context window) or disappear (when context is lost). There is no middle ground where the pact adapts to changing conditions.
Why This Breaks Long Tasks
Short tasks fit within a single context window. Long tasks do not. For any task that takes more than a few hours:
- The original pact gets pushed out of context
- New sessions start without the coordination agreement
- Agents revert to their default behavior
- Conflicts that the pact was designed to prevent start occurring
The Session Boundary Problem
Every agent session is a clean slate. The agent loads its system prompt, reads any persistent memory files, and starts fresh. If the pact was only verbal (established in the chat), it is lost. If it was written to a file, the agent needs to know to read that file and treat it as binding.
What Would Fix This
Agent pacts need three properties they currently lack:
Persistence - Written to a durable store that survives session boundaries. Not just memory files but structured agreements with clear terms.
Expiration awareness - The pact should include "this agreement expires when X" so agents know when to renegotiate rather than silently dropping the agreement.
Renegotiation triggers - When conditions change (new information, failed assumptions, scope changes), agents should automatically revisit the pact rather than continuing with an outdated agreement or no agreement at all.
Until these are built into agent frameworks, long-running multi-agent tasks will keep breaking at session boundaries.
Fazm is an open source macOS AI agent. Open source on GitHub.