A structural representation that makes emergent LLM-agent collaboration explicit - tasks, environment, and agents on one record - without prescribing agent reasoning, behavior, or communication order. Infuse turns that record into a better collaboration policy.
A collaboration policy - an executable graph, or code and instructions - rarely controls everything. Agents deviate from workflows, orchestrators decompose work on the fly, and in open-ended settings agents organize collaboration themselves. Flexibility helps performance, but the structure that emerges is hard to track, and therefore hard to improve.
TEA couples an Agent Interaction Graph (asynchronous activations and events) to a Task Activity Graph (immutable task versions and the actions that create them) and the shared environment, attributing task changes to activations and evidence to exact versions. Merging activations into nodes and counting messages per edge recovers the topology; activation timing and message paths expose the execution rules, including the ones the policy never specified.
Infuse queries the record to check the recovered policy against task progress and environmental outcomes: which parts helped, which hurt, and which never showed an effect. Each part is kept, revised, or left open, and the result is an updated policy Γ′ for the next episode. Starting from an empty policy, one Infuse update improves task performance in both benchmarks, code refactoring and Silo-Bench. A separate agent writes Γ′ as lessons for the team and for each agent.
Task versions and the actions that create them. A version carries a persistent identity, a goal and evaluation rule, and a state. Versions are immutable, so branching and merging work keeps its full history.
Events and activations. An activation is one reasoning cycle of an agent: its interval, received events, and ordered tool calls with results. Events carry any payload and can wake other agents.
The world agents observe and change. Environment calls return fresh events into the AIG, so task declarations and evidence stay tied to externally verifiable effects.
OPEN(δ, σ)∅ → qCreate a fresh task identity with an initial version.
EDITδ(q, δ′)q → q′Revise the goal or evaluation rule while preserving the reported state.
UPDATEσ(q, σ′)q → q′Change the reported state while preserving the task specification.
SPLIT(q, {(kj, δj, σj)})q → {q1, …, qm}Decompose one version into versions with distinct task identities.
JOIN({qj}, k′, δ′, σ′){q1, …, qm} → q′Integrate versions from distinct identities into a single successor version.
CLOSE(k)propagate closureClosure propagates backward through recorded ancestry; a split's parent closes only when all of its children are closed.
Besides task and environment tools, TEA provides sendh(e, J), which delivers an event to agents
J (a new message or a forwarded event), and attachh(φ), which attaches
evidence φ = (h, x, q) to an exact task version. Evidence stays with that version and
does not carry over to later revisions. For a task version q, the attached evidence, its
providers, and its content together provide signals of task progress; interpretations from different agents may agree,
partially agree, or disagree, and different interpreters may weight sources differently.
| Tool call | AIG | TAG | Env. | Evidence |
|---|---|---|---|---|
| task tool u ∈ UT | D′ | T′ | Z | Φ |
| environment tool u ∈ UZ | D′ | T | Z′ | Φ |
| attachh(φ) | D′ | T | Z | Φ′ |
| sendh(e, J) | D′ | T | Z | Φ |
TEA transitions: highlighted cells are updated by the call.
A collaboration policy is an interaction topology plus coordination rules. Given a TEA record, aggregate the activations in a window by agent: agents become nodes, and message counts become edge weights. Activation timing and message paths then reveal execution order, parallel overlap, repetition, conditional routing, and joins.
Topology recovery. Take the activations that start and finish within a window [t, t′] and merge them by agent. The result is a weighted graph G[t,t′] = (V, E, w) whose edges connect agents that exchanged messages, weighted by message counts. The full topology is recovered once the records cover every edge the policy can create.
Execution rules. Consecutive activations whose starts fall within τ of each other and whose intervals overlap are grouped. For completed activations with start sh and finish fh: h ≺ h′ iff fh ≤ sh′ (execution order), and h ∥τ h′ iff |sh − sh′| ≤ τ (overlap). Along activation paths these give sequential, parallel, repeated, conditional, and join rules, including ones the policy never specified; a worked example is in the paper.
TEA is also the interface through which agents communicate, transform tasks, and act in the environment. Infuse takes the trace of an episode - starting from an empty policy - and proposes the next one.
DT ← fexecute(V, Γ)
The team V attempts the task under policy Γ, and TEA records the collaboration as it unfolds.
Γ̂ ← frecover(DT)
Aggregate activations by agent to recover the observed topology and infer candidate execution rules.
Γ′ ← fanalyze(Γ, Γ̂, DT)
Query the record to compare the recovered policy with the trace, assess coordination against task progress and environmental outcomes, and decide what to keep, add, revise, or leave open.
In the paper's implementation, a separate gpt-6-sol agent consolidates this guidance into global lessons for
the team and private lessons for individual agents, forming Γ′. Infuse itself is
agnostic to how Γ′ is produced.
Agents run on TEA asynchronously. In each activation an agent acts on its environment view, its task-graph view, received messages, and private memory, and every tool call is recorded. An agent ends its own activation after a round of reasoning, can sleep and reactivate after a delay, and is activated by an incoming message; a sender can also interrupt an active recipient, ending its activation and starting a new one with the message.
Two domains, five gpt-5.6-terra agents with no predefined roles: code refactoring (modularize a
redundant codebase; LoC reduction while passing fixed tests, six repositories, 100-step budget) and
Silo-Bench (30 distributed reasoning tasks where every agent must answer a query that needs the whole team's
information). A separate gpt-6-sol agent derives the Infuse guidance.
Access to TAG and to direct communication is ablated in four conditions.
Code (shop) | Silo-Bench | |||||
|---|---|---|---|---|---|---|
| Setting | Pass ↑ | Rall ↑ | Rpass ↑ | Success ↑ | S ↑ | P ↑ |
| TAG + comm | 4/5 | 46.9 | 43.0 | 0.77 | 0.81 | 0.85 |
| TAG only | 3/5 | 35.3 | 35.2 | 0.77 | 0.81 | 0.83 |
| Comm only | 3/5 | 20.2 | 19.5 | 0.60 | 0.73 | 0.78 |
| Neither | 4/5 | 21.3 | 20.8 | 0.03 | 0.15 | 0.20 |
Code refactoring on shop: five runs per condition; Pass counts runs passing all tests;
Rall and Rpass are mean LoC reductions (%)
over all and passing runs. Silo-Bench: 30 tasks; S is the fraction of exactly correct answers,
P the partial-credit score. Best values in bold.
shop: mean of five runs per condition, shading
±1 SD. The gap between the TAG conditions and the others widens during execution.
TEA's shared task graph outperforms direct communication alone on both benchmarks, while combining the two yields the strongest overall results.
TEA with and without one Infuse update, five runs per repository and condition.
Infuse enables TEA to refine its collaboration policy from experience and improve overall task performance.
Code refactoring before (TEA) and after one Infuse update (Inf.)
| Easy | Medium | Hard | Average | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Codebase | shop | plan | ledger | freight | stay | payroll | ||||||||
| TEA | Inf. | TEA | Inf. | TEA | Inf. | TEA | Inf. | TEA | Inf. | TEA | Inf. | TEA | Inf. | |
| Pass rate ↑ | 0.80 | 0.80 | 0.80 | 0.80 | 0.60 | 0.80 | 0.60 | 0.80 | 0.40 | 0.60 | 0.60 | 0.40 | 0.63 | 0.70 |
| LoC reduction (%) ↑ | 43.0±10.2 | 56.6±8.9 | 28.3±5.9 | 31.6±3.7 | 22.1±3.4 | 25.1±1.7 | 20.6±4.0 | 22.1±3.5 | 20.2±2.6 | 24.8±4.4 | 18.8±2.5 | 22.2±9.0 | 25.5±9.2 | 30.4±13.3 |
Average is the unweighted repository mean; ± denotes sample SD across passing runs, or across repositories for
Average. Better paired values in bold; ties unbolded. Pass rates rise on ledger, freight, and
stay, remain unchanged on shop and plan, and only fall on payroll.
Silo-Bench before (TEA) and after one Infuse update (Inf.), 30 tasks
| TAG + comm | TAG only | Comm only | Neither | |||||
|---|---|---|---|---|---|---|---|---|
| TEA | Inf. | TEA | Inf. | TEA | Inf. | TEA | Inf. | |
| Success ↑ | 0.77 | 0.83 | 0.77 | 0.73 | 0.60 | 0.67 | 0.03 | 0.17 |
| S ↑ | 0.81 | 0.86 | 0.81 | 0.75 | 0.73 | 0.79 | 0.15 | 0.19 |
| P ↑ | 0.85 | 0.88 | 0.83 | 0.78 | 0.78 | 0.82 | 0.20 | 0.24 |
Bold marks the best value in each row. Infuse raises success in three of four settings; TAG-only success slightly falls from 0.77 to 0.73.
send removed, there is no edge to draw.
